Self-Paced Convolutional Neural Network for PolSAR Images Classification

Abstract

Fully polarimetric synthetic aperture radar (PolSAR) can transmit and receive electromagnetic energy on four polarization channels (HH, HV, VH, VV). The data acquired from four channels have both similarities and complementarities. Utilizing the information between the four channels can considerably improve the performance of PolSAR image classification. Convolutional neural network can be used to extract the channel-spatial features of PolSAR images. Self-paced learning has been demonstrated to be instrumental in enhancing the learning robustness of convolutional neural network. In this paper, a novel classification method for PolSAR images using self-paced convolutional neural network (SPCNN) is proposed. In our method, each pixel is denoted by a 3-dimensional tensor block formed by its scattering intensity values on four channels, Pauli’s RGB values and its neighborhood information. Then, we train SPCNN to extract the channel-spatial features and obtain the classification results. Inspired by self-paced learning, SPCNN learns the easier samples first and gradually involves more difficult samples into the training process. This learning mechanism can make network converge to better values. The proposed method achieved state-of-the-art performances on four real PolSAR dataset.

SAR4LCZ-Net: A Complex-Valued Convolutional Neural Network for Local Climate Zones Classification Using Gaofen-3 Quad-Pol SAR Data

“…In order to classify all pixels in a patch of urban area simultaneously, a fixed-feature-size CNN was used in [17]. A deep self-paced CNN [18] was employed for river area classification, which retains mutual information between channels and texture information from the raw quad-pol SAR data. For building and vegetation classification [19], a quad-pol SAR image classification method based on residual network (ResNet) and deep autoencoder was proposed.…”

Self-paced hybrid dilated convolutional neural networks

“…In the network model, the size of the dilated convolution kernel in each dilated convolution layer is 3 * 3, and the stride in the convolution operation is all set to 1. The network structure only uses a small-sized kernel because it can not only greatly reduce the parameters in the network structure, but also obtain better results [5]. The first dilated convolutional layer kernel size is 32 * 3 * 3 * d, where d is the number of channels of the input image.…”

“…There are three parameters to be optimized in (5), namely W, b and v. We use the control variable method to optimize the three parameters and obtain better network model parameters finally. The algorithm of SPHDCNN is shown in Algorithm 2.…”

“…In order to analyze the effect of introducing hybrid dilated convolution in the model and the impact of different dilated rate sequences, we tested the performance of different dilated rate sequences with dilated rate in the range of (1,2,3,5). In this experiment, we use four layers of convolutions, and each layer uses 32 convolution kernels of size 3 * 3.…”